LFM3 Guide: Liquid AI's New On-Device Models Explained

Last updated: October 30, 2026

LFM3 is Liquid AI’s new family of open-weight models, released October 27, 2026. It puts text, vision, audio, omni and decision models on one 1.7B backbone. The text model runs in about 1 GB at 4-bit, has a 256K context window, and is trained for local agents and tool calling.

This guide is based on Liquid AI’s LFM3 release post on Hugging Face, the LFM3-1.7B model card and Ollama’s lfm3 model page. All benchmark figures below are Liquid AI’s own and have not been independently verified.

Contents

What’s new in LFM3
The six LFM3 models
Architecture changes
LFM3-1.7B for local agents
Vision, audio and omni models
d1 decision models
How to run LFM3
Limitations and licensing
Frequently asked questions

What’s new in LFM3

Liquid AI describes four main changes compared with the earlier LFM2 family:

More capability per parameter: LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters.
Every modality on one backbone: six 2B-class models trained from LFM3-1.7B-Base, including Liquid’s first omni model and the first LFM vision model that accepts video.
Built for the edge: new in-house vision and audio encoders and a text model that fits in roughly 1 GB at 4-bit.
Made to be specialized: every model can be fine-tuned on a single GPU.

For power users, that means small models you can run offline, fine-tune cheaply and mix by task.

The six LFM3 models

Model	Input	Output	Best for
LFM3-1.7B	Text	Text	Local agents, tool calling, structured output
LFM3-VL-2B	Text, images, video	Text	On-device visual understanding, robotics
LFM3-Audio-2B	Text, audio	Text, speech	Voice chat, speech recognition, speech synthesis, translation
LFM3-Omni-2B	Text, images, video, audio	Text, speech	One model that sees, listens and talks back
d1-2B	Text, images, video	Calibrated answers	Fast classification and structured decisions
d1-omni-2B	Text, images, video, audio	Calibrated answers	Decisions when inputs include audio

Architecture changes

According to the release post, LFM3-1.7B-Base has 24 layers in a “tri-hybrid” mix: 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. Other changes include head-wise gated attention, partial RoPE on the global layers and a sandwich norm around each block.

The base model was pre-trained on 22 trillion tokens with a 128,000-token vocabulary. The final 3.5 trillion tokens were a mid-training stage with more agentic traces and tool calls, which also extended the context window to 256K tokens. Liquid AI reports that with 8K tokens in context, LFM3-1.7B runs about 25% faster per parameter than LFM2-1.2B on edge devices.

LFM3-1.7B for local agents

LFM3-1.7B is post-trained to run inside agent harnesses such as Hermes Agent, Pi and OpenClaw. Liquid AI’s largest reported gain is on WebArena-Verified, where it scores 38.42 against 3.02 for LFM2.5-2.6B. The model card also lists 45.60 on LiveClaw, ahead of Qwen3.5-9B at 35.77.

Settings worth knowing from the model card:

Sampling: for chat, tool use and agents use temperature=0.1, top_k=50, repetition_penalty=1.1. For math and science use temperature=1.0, top_p=0.95, top_k=20.
Reasoning: thinking is on by default. You can pass enable_thinking=False, but Liquid AI says this can be inconsistent on reasoning-heavy tasks. Earlier reasoning is dropped unless you set preserve_thinking=True.
Tool calls: the model emits XML tool calls, and function.arguments must be a Python mapping, so decode JSON strings first.
Tool search: mark tools with defer_loading=True to keep large tool catalogs out of the prompt until a search tool returns them.

If you prefer a larger open-weight model for harder reasoning, compare our Mistral Large 4 guide.

Vision, audio and omni models

LFM3-VL-2B uses a new 164M-parameter vision encoder that Liquid AI says needs 3x less compute than the SigLIP2 encoder in earlier LFM vision models. With 1.9B parameters against 3.1B for LFM2.5-VL-3B, it raises the average across 62 vision benchmarks from 57.6 to 71.4, and it is the first LFM vision model to accept video.

LFM3-Audio-2B handles voice chat, speech recognition, speech synthesis and translation across 18 languages, and can turn spoken requests into tool calls. Its causal audio encoder can prefill while you speak, with time to first token as low as 50 ms. The new LiquidCodec tokenizer generates speech at about 87x real time on a single CPU thread on an Apple M4 Max. For hosted speech options, see our Voxtral 3 speech models guide and Grok Voice 3.0 guide.

LFM3-Omni-2B takes text, images, video and audio and replies in text or speech, so one model can replace a separate speech recognizer, vision model, language model and voice. Liquid AI says it beats models up to 10.7B parameters, such as Qwen2.5-Omni-7B, on Daily-Omni, WorldSense and AV-SpeakerBench.

d1 decision models

The d1 models answer a list of questions about a situation, such as a support ticket, invoice, app screen, photo or voice note, in a single forward pass, and each answer comes with a calibrated probability. Beyond yes/no, choices and scores, they can now point at objects, read values from documents, estimate numbers with a range and select every option that applies.

d1-2B replaces d1-3B with a third fewer parameters and scores 75.5 against 64.0 on Liquid AI’s ImajevBench. Choose d1-omni-2B when inputs include audio. Our Perplexity Decisions API guide covers a hosted alternative, and the Ollama new features guide explains Ollama’s --ask flag for decision models.

How to run LFM3

Pick the model for your inputs from the table above, and the checkpoint for your hardware: GGUF for llama.cpp, Ollama and LM Studio; MLX for Apple Silicon; BF16 for vLLM, SGLang or Transformers; FP8 or NVFP4 for SGLang on GPUs; or ONNX.
For the quickest local test, install Ollama 0.40.12 or later and run ollama run lfm3.
To drive a local agent, run ollama launch hermes --model lfm3 or ollama launch openclaw --model lfm3, as listed on Ollama’s model page.
In Python, run pip install -U transformers and load LiquidAI/LFM3-1.7B with a text-generation pipeline and trust_remote_code=True.
For the multimodal models, use transformers>=5.21.0 for LFM3-VL-2B, transformers>=5.20.2 for the audio and omni models, and transformers>=5.19 for d1-2B.
Apply the recommended sampling settings, then fine-tune on your own data with Liquid AI’s SFT, DPO, GRPO or continued pre-training notebooks, each sized for a single GPU.

Ollama added the LFM3 family in version 0.40.12, and its page shows the text model at 1.1 GB to 1.7 GB with a 256K context window and tool and thinking support.

Limitations and licensing

The model card says LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning, and suggests retrieval or search for knowledge-heavy tasks.
Turning off thinking with enable_thinking=False can behave inconsistently.
The text model supports 16 languages, listed on its model card; the audio model covers 18 languages.
The weights use the LFM Open License v1.0. Check its terms before commercial use.
All benchmark results quoted here are published by Liquid AI.

Frequently asked questions

What is LFM3?

LFM3 is Liquid AI’s third generation of Liquid Foundation Models, released on October 27, 2026. It is a family of open-weight models in the 2B parameter range, all built on one backbone: a 1.7B text model, a vision model, an audio model, an omni model and two d1 decision models.

How big is LFM3-1.7B and what is its context window?

LFM3-1.7B has 1.73 billion parameters and a 256K-token context window covering input and output. Liquid AI says it needs about 1 GB of memory at 4-bit precision, so it fits on phones and laptops. Ollama lists its downloads at 1.1 GB to 1.7 GB.

Can I run LFM3 in Ollama or LM Studio?

Yes. The LFM3-1.7B model card lists GGUF checkpoints for llama.cpp, Ollama and LM Studio, plus MLX, ONNX, BF16, FP8 and NVFP4 versions. In Ollama, run ollama run lfm3. Ollama added the LFM3 family in version 0.40.12.

Is LFM3 free for commercial use?

The LFM3 weights are open and published on Hugging Face under the LFM Open License v1.0. Read that license on the model card before commercial deployment, because its terms, not the word open-weight, decide what you are allowed to do.

Is LFM3 good for coding?

Not as a primary coding model. The LFM3-1.7B model card says it is not recommended for agentic coding or complex multi-step reasoning because of its small size. It is aimed at local agents, tool calling, structured outputs and fine-tuning for a specific domain.

Sources

Image: Liquid AI logo by Human Agency, Inc., public domain (trademarked), via Wikimedia Commons. Placed on a white background and resized. Liquid AI is a trademark of its owner.

LFM3: Power-User Guide to Liquid AI’s On-Device Text, Vision, Audio and Decision Models